The Twenty-Seven

Headlines last week said the same thing: Adopt AI, and your headcount grows.

For example the FT, who seem to be slipping in the quality of their writing, I'm sad to see, published it as "Heavy corporate AI spenders add staff faster than peers". This is simply not accurate. It stems from reading a lazy summary, and not the underlying white paper.

Twenty-one thousand companies in one study. Two thousand two hundred executives across twenty-one countries in the other. Real data, careful method, no obvious axe.

Then, yesterday, Demis Hassabis published a warning that we are a few short years from AGI and urgently need regulation before we arrive.

Both cannot be true. So I read the underlying papers rather than the coverage.

One of the studies was enabled by a credit card company.

The other was enabled by a landlord.

Ramp is a corporate spend platform. Its product is issued to employees, and its business grows as its customers' payrolls grow. On 30 June, with Revelio Labs, it published a study of 21,559 firms finding that companies making serious AI investments grew headcount by around 10% over two years, with entry-level headcount up 12%.

JLL is one of the largest commercial real estate firms on earth: $26.1 billion of revenue. Its business grows as companies occupy more square footage, which they do by employing more people to sit in it. In June it published its Future of Work Survey, reporting that most organisations anticipate net headcount expansion, with AI redesigning roles rather than removing them.

I am not saying either organisation fabricated anything. I've read both properly, and the accusation would be false.

The Ramp paper is the best firm-level work on this question that exists, because it doesn't ask anyone anything. No exposure indices, no executives guessing. It watches actual money move to actual AI vendors and links it to actual workforce records.

Ramp splits adopters by AI spend per employee. The low-intensity group averages $2.78 per employee per month, and across every outcome the paper measures, headcount, entry-level, engineering, sales, marketing, admin, customer service, that group's result is statistically indistinguishable from zero. The authors say it plainly: chat subscriptions are not enough, and neither is a few months of experimental spend.

One of the paper's own authors, Ryan Stevens, published separate work this year finding direct substitution of labour for AI in Ramp's own data: firms cutting freelance spend as AI spend rose. Same company. Same dataset. Opposite conclusion. It received no coverage whatsoever.

JLL is equally straightforward. But unlike the Ramp report, it never claims to have measured what happened, instead it is a survey of what 2,200 executives expect. Opinion, honestly labelled.

So nobody lied.

And yet the week produced a headline that cannot be true.

In the Ramp study, 'Adoption intensity' is measured as AI spend per employee. So the denominator does the work.

Mean headcount of the high-intensity group (the firms that grew 10%, the entire basis of the reassurance now circulating) is 26.7 people. Low-intensity adopters average 193.4. Never-adopters, 103.9.

Multiply $33.67 per employee per month by 26.7 and you get a company spending roughly nine hundred dollars a month on AI. That is not organisational transformation. That would not cover a day of a mid-tier agency's project rate.

Stated honestly: small, tech-adjacent startups (34% venture-backed against 9.5% of never-adopters), already growing three times faster before a single dollar reached a single model, spend a lot per head on AI, and also hire people. That isn't AI creating jobs. That's a Series A creating jobs. By month 24 the estimate reaches 57% headcount growth attributed to AI adoption, a figure nobody, including the authors, believes describes the causal effect of a $900 monthly spend.

In short, what is being reported is (at best) correlation, not causation, and therefore it is lazy and inaccurate. It should not give the comfort that is intended.

Put the two reports side by side, and watch them make the same omission in their summaries and the reporting on them: more evidence that the bottom rungs of the ladder are missing or broken.

Ramp's Table 4 measures workforce composition rather than firm size. Among low-intensity adopters (mean headcount 193, companies still smaller probably than the one of you work for) entry-level share fell 0.52 percentage points while manager-plus share rose 0.66. Quietly, the large adopters are tilting away from juniors while twenty-seven-person startups pull the average the other way. Its own literature review cites the ADP payroll finding of a roughly 16% employment decline for 22-to-25-year-olds in the highest-exposure occupations. It cites it. It does not refute it.

Now JLL. Under the heading describing a broadly positive outlook, there is a chart. Workforce expansion beats workforce reduction, 60% to 39%. That's the finding that travelled.

Look four rows down. Reduced entry-level hiring: 49%. Large-scale entry-level hiring: 49%.

A dead heat, filed under "diverging approaches in the market" and reported as no clear trend. Half the C-suite of the developed world expects to hire fewer graduates and the other half expects to hire more, and that is the most consequential split in the survey.

In March, in The Power of And, I wrote about an Anthropic study showing no systematic rise in unemployment in the most AI-exposed occupations, but this was alongside a 14% slowdown in the hiring of younger workers into those same occupations. The jobs weren't disappearing. The pipeline into them was already closing.

Four months on, two more reports, two more reassuring summaries, and in both the entry-level number is the one that didn't make the top line. That is not a conspiracy. Nobody coordinated it. It is simply what happens when the party writing the summary is paid by the headcount.

Which brings me back to Hassabis. He isn't more useful because he's more right, but because he's asking a better question and is honest that he doesn't know the answer. He has an interest too: his proposed Standards Body would be largely industry-funded, its benchmarks set in consultation with the frontier labs, startups exempt. A licensing regime you help write is a licensing regime you win. That makes him interested, not wrong.

Two studies. One from a company paid per employee, one from a company paid per desk. Both careful. Both caveated. Both converted within hours into the most comfortable headline of the year by people who didn't read past the abstract, for an audience that didn't want them to.

Comfort is the tell. When research tells you the most disruptive technology of your working life will cost you nothing, the first question is not is this true. It is who benefits from me believing it followed immediately by what's in the appendix.

We are not fine. We are unmeasured, by instruments belonging to people who would prefer a particular answer. And the reassurance of an entire white-collar generation currently rests on twenty-seven people and a nine-hundred-dollar invoice.

Brandflow is written by Justin Billingsley, who has spent his career on all three sides of the industry's table: senior client, global agency leader, technology founder. First published 16 July 2026 in the Brandflow newsletter on LinkedIn.